llama : only copy used KV cache in get / set state #1272
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
SlyEcho
reviewed
May 2, 2023
llama.cpp
Outdated
Comment on lines
2483
to
2506
| { | ||
| // copy k: k layout is n_layer > n_ctx (tokens) > n_embd | ||
| const uint8_t * k_data = (uint8_t *) kv_self.k->data; | ||
| const size_t elt_size = ggml_element_size(kv_self.k); | ||
|
|
||
| for (int il = 0; il < n_layer; il++) { | ||
| const size_t offset = il * n_ctx * n_embd * elt_size; | ||
| const size_t size = kv_ntok * n_embd * elt_size; | ||
| memcpy(out, k_data + offset, size); out += size; | ||
| } | ||
| } | ||
|
|
||
| { | ||
| // copy v: v layout is n_layer > n_embd > n_ctx (tokens) | ||
| const uint8_t * v_data = (uint8_t *) kv_self.v->data; | ||
| const size_t elt_size = ggml_element_size(kv_self.v); | ||
| const int n_layer_embd = n_layer * n_embd; | ||
|
|
||
| for (int ile = 0; ile < n_layer_embd; ile++) { | ||
| const size_t offset = ile * n_ctx * elt_size; | ||
| const size_t size = kv_ntok * elt_size; | ||
| memcpy(out, v_data + offset, size); out += size; | ||
| } | ||
| } |
There was a problem hiding this comment.
Instead of writing the tensor copy code manually, why not use ggml?
I proposed something like this before:
{
const size_t elt_size = ggml_element_size(kv_self.k);
char buffer[4096]; // should be enough
ggml_context *cpy_ctx = ggml_init({ sizeof(buffer), buffer, true });
ggml_cgraph gf{};
gf.n_threads = 1;
ggml_tensor * kout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_ntok, n_layer);
kout3d->data = out;
out += ggml_nbytes(kout3d);
ggml_tensor * vout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.v->type, kv_ntok, n_embd, n_layer);
vout3d->data = out;
out += ggml_nbytes(vout3d);
ggml_tensor * k3d = ggml_view_3d(cpy_ctx, kv_self.k, n_embd, kv_ntok, n_layer, elt_size*n_embd, elt_size*n_embd*n_ctx, 0);
ggml_tensor * v3d = ggml_view_3d(cpy_ctx, kv_self.v, kv_ntok, n_embd, n_layer, elt_size*n_ctx, elt_size*n_ctx*n_embd, 0);
ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, k3d, kout3d));
ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, v3d, vout3d));
ggml_graph_compute(cpy_ctx, &gf);
}
|
Thanks! Will do. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Per comments in #1169 and #1247, reduces the size of serialized model state and session files by only storing the used portions of the KV cache. This corresponds to the tokens evaluated so far in the session, so the size scales with # of tokens, up to max state size at full context.
Changes
llama_copy_state_dataandllama_set_state_datanow use the number of evaluated tokens (as inllama_get_kv_cache_token_count) to (de)compact the KV cache data. As a result, whilellama_get_state_sizeis unchanged, its semantics differ: it is now the maximum buffer size required by the get / set state APITesting
chat-13Bon 30B and 65B. Ran cold and warm and ensured same output.examples/save-load-stateResults
Original sizes:
New sizes: